Learning to Ask Critical Questions for Assisting Product Search

Zixuan Li, Lizi Liao, Tat-Seng ChuaarXiv 2024 · DOI · arXiv

Product search in eCommerce often feels like a game of "hot or cold." Users frequently start with vague queries, leaving systems to either guess intent from a trail of clicks (session-aware recommendation) or bombard the user with clarifying questions too early (conversational search). A new framework, DualSI, bridges this gap by waiting for the user to explore, then proactively asking "critical" questions to zero in on the target.

The Challenge of Intent Clarification

Current product search systems face a fundamental mismatch in how they handle user intent:

  1. Vague Queries: Initial search terms are often too general to pinpoint a specific product.
  2. The Guessing Problem: Session-aware models treat clicks as implicit feedback but are essentially guessing preferences based on sequences.
  3. The Interrogation Problem: Conversational models often assume a clear target from the start and require intensive user involvement, which can be inefficient during early browsing.

DualSI identifies the "sweet spot": allowing users to browse first to clarify their own intent, then intervening with the right questions at the right time.

The DualSI Framework

The core of DualSI is a dual-learning architecture that combines implicit session feedback with proactive, targeted questioning.

The DualSI framework: Selection Net identifies critical questions, while Ranking Net refines product ranking based on user feedback.

1. Contextualized Selection Net

This module decides what to ask by processing the user's click-stream and search query. It uses Transformer blocks to capture complex, non-linear transition patterns in the click sequence.

The system's key innovation is the Utility Score ($\alpha$). The model calculates an attention-based utility score to measure how much "useful information" a clicked item (or a "synthetic item" composed of attribute combinations) provides for finding the target. The attributes of the highest-scoring synthetic item become the "critical question" presented to the user.

The Selection Net architecture utilizes Transformer blocks and an attention mechanism to calculate utility scores for item attributes.

2. Dual Ranking Net

Once the user answers (e.g., "Yes, I want a waterproof watch"), the Ranking Net takes over. It reuses the frozen encoder from the Selection Net and combines session history, the original query, and the new explicit feedback to rank the entire product pool, aiming to place the true target item at the top.

Experimental Results

The authors evaluated DualSI using the Diginetica dataset. The model significantly outperformed both session-only and interactive baselines across standard metrics like Mean Reciprocal Rank (MRR) and NDCG.

Model MRR (1 Question) Top-3 Accuracy (1 Question)
QSBPS (Interactive Baseline) 0.1597 0.1617
SCEM (Session-only Baseline) 0.1619 0.1868
DualSI (Proposed) 0.2331 0.2575

Key Insights: * Significant Gains: DualSI achieved a 45.96% MRR improvement over the interactive baseline when asking just one question. * Efficiency: The model maintained high performance even when questions were asked very early in the session (e.g., after only 1-3 clicks), suggesting it can effectively shorten the user's path to purchase. * The Utility of a Click: By explicitly modeling "utility," the system can ignore exploratory noise and focus on attributes that are highly indicative of the final goal.

Conclusion and Future Work

DualSI demonstrates that proactive questioning, when timed correctly and targeted at high-utility attributes, can drastically improve product search efficiency. Future research directions include better utilization of negative feedback (rejected attributes) and developing mechanisms to dynamically decide when the system has enough confidence to prompt the user.

Summary updated

Abstract

Product search plays an essential role in eCommerce. It was treated as a special type of information retrieval problem. Most existing works make use of historical data to improve the search performance, which do not take the opportunity to ask for user's current interest directly. Some session-aware methods take the user's clicks within the session as implicit feedback, but it is still just a guess on user's preference. To address this problem, recent conversational or question-based search models interact with users directly for understanding the user's interest explicitly. However, most users do not have a clear picture on what to buy at the initial stage. Asking critical attributes that the user is looking for after they explored for a while should be a more efficient way to help them searching for the target items. In this paper, we propose a dual-learning model that hybrids the best from both implicit session feedback and proactively clarifying with users on the most critical questions. We first establish a novel utility score to measure whether a clicked item provides useful information for finding the target. Then we develop the dual Selection Net and Ranking Net for choosing the critical questions and ranking the items. It innovatively links traditional click-stream data and text-based questions together. To verify our proposal, we did extensive experiments on a public dataset, and our model largely outperformed other state-of-the-art methods.
BibTeX
@misc{li2024learning,
  author = {Zixuan Li and Lizi Liao and Tat-Seng Chua},
  title = {{Learning to Ask Critical Questions for Assisting Product Search}},
  year = {2024},
  eprint = {2403.02754},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.02754},
}
Figure